Team DataFam Tableau User Group 16 Oct 2026
"Mastering AI Narrative Generation via Tableau Extensions" by David Nguyen
1h 27m 14,691 words transcribed
Summary
Team DataFam TUG meeting #94. David Nguyen, a four-time Tableau DataDev Ambassador, opens a three-part series on generative AI and Tableau extensions with a session on AI narrative generation. After Dreamforce '26 put Tableau Studio, Tableau Knowledge, Proactive Intelligence and Agent in Dashboards on the roadmap, he asks why anyone would still build their own AI extension, and answers with a live demo: a 'last-mile narrator' extension on a Superstore dashboard that reads exactly what the viewer sees and writes a summary for the CEO, an anomaly and root-cause analysis for an analyst, or a different narrative per zone, with the model of your choice. He covers how dashboard extensions work (the .trex manifest, initializeAsync, reading summary data with filters and row-level security already applied, the Settings API), building one in minutes with an AI coding agent, a FastAPI backend in a Docker container on Azure, benchmarking three models with promptfoo, CI/CD with GitHub Actions, six layers of security, and what one insight actually costs in tokens. Chris McClellan hosts the Q&A on governance, data residency and cost, and runs a quiz David wrote.
Listen to this TUG
Prefer to listen? This meeting is also an episode of the Tableau TUG recordings podcast.
Key points
- A custom GenAI extension is the last layer on top of what you already built: it reads exactly what the viewer sees and writes the story with the model you choose, with no rebuild and no migration.
- If Tableau's own agent and Q&A already answer your question, use them; a custom extension means source code to maintain and a backend to build and host.
- An extension reading summary data gets only what the viewer is supposed to see: filters and row-level security are already applied.
- Call tableau.extensions.initializeAsync() before any other Extensions API method; it is the handshake with the host.
- Send data to the model as CSV rather than JSON, which repeats every property name on every row and costs more tokens.
- Keep API keys on the server side, never in the front end or in workbook settings; AI coding agents often get this wrong.
- Lock the API to the extension's origin, cap spend and rate limits, and sanitize model output (for example with DOMPurify) before rendering it.
- Different models can all give correct answers yet tell different stories, so benchmark them side by side before switching.
- Output tokens cost about five times as much as input tokens; one insight is around two cents, but 100,000 insights is roughly $2,000.
- Deploying the model to a single region, such as Australia East, keeps the data in that country; a global endpoint may not.
Jump to a chapter (28)
- 0:00 Welcome to meeting #94
- 0:42 Introducing David Nguyen
- 3:28 Agenda
- 6:25 September 2026: Tableau went agentic
- 8:12 Demo: an AI narrator on a Superstore dashboard
- 12:30 The board meeting: five readers, five answers
- 17:18 Where the GenAI extension fits
- 19:15 Native, custom or MCP?
- 22:10 What the narrator picked up
- 23:00 Several narrators, one dashboard
- 27:17 Tableau extensions 101
- 30:10 How the project is structured
- 35:45 How an extension works, in five steps
- 38:40 The Settings API
- 40:00 Summary data and saving tokens
- 41:43 Building an extension with an AI coding agent
- 47:30 Inside the backend
- 50:07 What agents get wrong
- 53:10 Insight modes and prompt templates
- 54:17 Structured output and business definitions
- 55:51 Benchmarking three models
- 59:55 CI/CD with GitHub Actions
- 1:00:57 Securing a Tableau extension
- 1:03:42 How tokens work and what an insight costs
- 1:08:30 Key takeaways and resources
- 1:10:36 Q&A: governance, data residency and cost
- 1:20:33 Survey and quiz
- 1:25:30 Next meeting and wrap-up
Transcript
0:01 Chris McClellan: So welcome to Team Data Fam Tableau User Group meeting number ninety-four, almost at a hundred. Uh, this TUG aims to capture all the in-person and unrecorded presentations and give them a wider audience. So I don't really care where you've presented before. I don't actually care if you've never presented before.
0:18 Everyone's welcome here. We'll find a slot for you. Email me at team.DataFam@Gmail.com, and we'll, um, find a spot.
0:27 But today we've got David talking about mastering AI narrative generation via Tableau extensions.
0:35 Sounds really cool. He's the expert. I'll hand it over to David, and he can take it from here.
0:42 David Nguyen: Thank you, Chris.
0:44 Uh, so for, uh, just a quick introduction, um, about, uh, me. Uh, so my name is, uh, David Nguyen.
0:51 Uh, so today's session, it should be the first session of the three topics that I, I have been working on.
0:58 Um, so I think the timing is just, uh, perfect because I think last month at the Dreamforce, right, uh, Tableau already went, uh, all-in on the agent test. They're talking about, like, Tableau Studio, Tableau Knowledge, Proactive Intelligence.
1:13 Uh, so I think it, it, it kind of a fair question here we should ask, like, if Tableau already built AI in, so why would anyone still build our own AI extension? In the next, uh, sixty minutes, um, we will talk about how you can leverage, uh, Tableau extension. Uh, we talk a little bit about AI, uh, how to build the extension, and how the custom extension here fit into, uh, the big kind AI ecosystem.
1:43 Um, yeah, so before we get start, uh, let me…
1:47 Just, like, very quickly about me.
1:50 Uh, so I'm the t- four-time, uh, Tableau Data Ambassador, so I got around, like, fifteen year of experience in, uh, software development, data analytics across different industry, in reinsurance, banking, investment, and healthcare. And, uh, recently I explore a few, a few thing, the agentic AI, immersive analytics in VR and SR. So I did a demo, uh, last month, uh, at the Sydney Tableau User Group about how to, uh, do the spatial analytics in, in VR.
2:20 So we have one session for that in, uh, December. So remember to check it out.
2:24 Um, yeah, so, um, just a few quick disclaimer before, um, anything go live.
2:33 Uh, so for, for the demo today, uh, so I'm, I'm using the sample data only. Uh, so basically that is Superstore, um, uh, data that everyone I think should know about that.
2:44 Um, and the second thing is the data that the AI generate, so it going to, uh, go out of Tableau. So if you want to, uh, create something like this, uh, you know, for your company, you should check with your IT, security, data governance, um, before you bring that into production.
3:04 Uh, we will talk more about that later. Uh, but, uh, through this, uh, presentation, I'm going to give you kind of the, the quick overview, um, about the process, the development from the beginning, uh, to how to debug the thing locally and then how to deploy that, uh, on production. Um, yeah.
3:24 Um, so, uh, let's, um…
3:28 So here the quick kind of, uh, agenda. I will try to keep it, like, within, um, sixty minutes. But in case run out of time, uh, just let me know, Chris.
3:38 Yeah. Uh, so i-in the first ten minutes, um, we will, um, talk about… Oh, sorry.
3:46 We will talk about where it fits, about the, the Dreamforce announcement with all the AI stuff, and then I will show a quick demo of the data story that, uh, uh, I, I create.
3:58 Uh, the second point is about the extension 101. So if you have been using extension before, I think that's fine.
4:05 Uh, if you haven't used that before, I think that's okay. We will cover what is it, uh, how it work, the structure of that, um, and how you can use AI agent to create one.
4:18 I will show you a quick demo of that.
4:20 Within five minutes, you can create your own, uh, extension. It's working.
4:25 Um, and the next one, it'll build with the agent, so that will be the live demo. Um, I, I already have captured the, the video as well, uh, so I'll show you how it work. And it got all the, the prompt as well, so you can try this out, uh, after the, the session.
4:41 Uh, next is about the back end and the AI inside.
4:46 So I will cover a little bit about, um, the back end, how it work, um, and then how you can do the benchmarking as well, uh, uh, based on the result, the output from, uh, different model.
4:59 Uh, number five, I will talk a little bit about CI/CD.
5:03 Um, it's continuous improvement, continuous deployment. Uh, I will give, uh, as, uh, example, but it will work in the same way on AWS or Google, uh, from, uh, how you can, uh, structure your project, um, how to use the GitHub, uh, for, uh, deployment, uh, how to deploy the front end, how to de-deploy the back end, uh, to a Docker container, uh, and a few tip and trick on how to, uh, bring that from local development to, uh, production.
5:36 Uh, and we'll cover a little bit about the security. So I will provide a few kind of tip and trick on how you can, uh, secure, uh, your extension.
5:46 And the next one will be about the token, how it work.
5:49 Uh, so let's say you already have the dashboard. You have the AI narrative, uh, built in.
5:55 Um, you will have like hundred user using that. H-how you are going to measure the cost?
6:01 Um, so that is so important.
6:03 Uh, and the last one will be the proper Q&A. So in case we run out of time, uh, I think I can take that offline, um, probably mean or you can just, uh, drop me a message.
6:13 Yeah.
6:15 Uh, so I think that concludes that our journey, uh, today.
6:21 So let me, uh, move to the next slide.
6:25 Yeah.
6:26 So let's start with, um, this slide. Uh, so Tableau, I think last, last one, if you want the, the, the keynote search when, um, owing on the agent test, they talk about Tableau Studio, Tableau Knowledge, Proactive Intelligence, Tableau Data App, Agent in Dashboard, so many things.
6:44 Um, so, so of the, the feature is still in beta, uh, so you can check it out, um, uh, on the new way to be available. And I, I believe there are a few sessions about, I think, Tableau Studio, I think next month, I think.
6:59 Uh, so remember to register for that. Uh, so here just a quick overview. So Tableau Studio, so that is the agent-based authoring workspace, workspace. So you can build the dashboard, you can build the data app by with the prompt, or you can drag and drop. So it's just kind of the evolution of the desktop. We have the Tableau Knowledge here, so that is the, the complex layer.
7:18 And we have a Proactive Intelligence. So I think that there'll be a lot of demo about it on how you can, um, integrate, uh, Tableau in the Slack, where it can just push the insight into Slack directly.
7:32 And, uh, we have the Tableau Data App as well.
7:35 Um, and Agent in Dashboard, uh, so I did a quick demo for that. You can check it out for some video, video demo on how to use that, uh, for the Q&A above this.
7:47 Um, and it is still in the beta, bring your own LLM.
7:51 Uh, so I think I'm not sure when it will be available, but this will be an option for you where you can bring in your LLM, uh, instead of using the one by, by, by Tableau.
8:03 Um, cool. Um, any, any question, uh, so far? Yeah.
8:08 Chris McClellan: No questions that I've seen so far.
8:10 David Nguyen: Great. Great. Great.
8:12 Cool. So here just a, a quick look of, um, the, the, the AI tension, uh, that, um, we are going to, um, explore, uh, today.
8:23 So just a, just a quick thing about the setup. Uh, so, uh, I'm using the same board dashboard, um, uh, the super admin, uh, store for demo purpose.
8:34 Um, so I think all but already familiar with the, uh, the data behind here.
8:39 Uh, so this one, it has like a few, a few dashboard, like the overview, product, uh, customer, uh, dimension here and there.
8:48 Uh, so, uh, the question here, um, so let's say I'm a CEO.
8:54 Um, I have the meeting in ten minute.
8:57 Uh, can you just build like a quick narrative for me for the meeting?
9:02 So the thing is, we have a nice dashboard. We have a lot of charts, but we need to have someone kind of to interpret or explain what is in the chart, but for different audience.
9:17 The CEO, he might have…
9:21 If you want, uh, let me open this, uh, go back to this. Yeah. So for the narrative, it will have the ability for you to generate different insight for different target audience.
9:34 So you have the ability to customize that.
9:38 If you want to have just a headline, the five-sentence headline, and the model already pre-in here. I click generate. That's it. If you want more about like data quality, see the data quality.
9:52 I will explain how it work at the back end. So basically, it's just kind of the custom prompt, um, that you need to, uh, specify.
10:00 Um, cool. Uh, I, I will go through the demo, uh, uh, for this one later. So basically, how it's working, you build your own extension, you add it to different places of the dashboard.
10:14 And this one, it will ha- You can configure setting for that as well, so it can answer different question.
10:21 It can answer an executive question.
10:24 Uh, let me, uh, open this one. Give me one second. I have the…
10:30 Yep. So here just kind of an overview of how, how it look like, and then we, we can, um, deep dive into that, um, later.
10:39 So let's say I have a overview dashboard here. I got the, uh, I think profits ratio, order the, by the sale across all the region in US. And then we got the monthly sale by segment here for consumer, corporate, home office.
10:57 And now, uh, let's say I'm not the executive, um, I'm an analyst. I want to find out more about the anomaly and the root cause, what happened in this area.
11:08 So we have like different modes here.
11:11 Insight mode, executive, we just provide kind of a summary in one or two line.
11:16 Anomaly, we will, we will deep dive a bit more insight.
11:21 Here just an example on, um, the output from, uh, for this, uh, demo. I configed, uh, three model, which will be Phi Mini, Phi-Voiceless Luna, and Phi-Voiceless Song.
11:35 Um, you can bring in any model that, you know, like Clockwork or, or Opus, Vapor, whatever that into this. It's very quick and, and easy. Just generate the key. I will cover, uh, more about that later.
11:50 Yeah.
11:51 Um, yeah, so here just the sim- session more finding from, um, the insight.
11:58 Uh, you can see like, uh, overall, I have how many. Okay, but even sale in this month, two point three million.
12:04 That is the profits. However, there's several location Have a negative margin. We can see Ohio in the red color here. I think, where Ohio? I think should be here. Yeah.
12:16 Negative twenty-one point seven percent. And then Tennessee also. Texas also minus fifteen percent.
12:23 Yeah. So that is kind of the, the output from the, the generative AI extension, uh, that bringing.
12:30 Yeah. So here is just kind of the data story that, um, I've been talking about.
12:35 And the CEO, he asked the question, "Where do we stand?" So the, the AI narrative, it will, um, generate an insight.
12:45 In the summary, that is the profit.
12:48 Okay, two point two K.
12:50 Healthy gross margin.
12:52 Why the, the best? That the Ohio trend is the biggest outlier at twenty-one point seven percent loss. So that is kind of the, the summary mode for the CEO.
13:04 And the category manager, so you can change to a different mode, to anomaly mode. Like, they want to know more about that. So what is really selling?
13:14 Uh, we will cover more about it in the act number two.
13:18 Um, and then we can move to the number three about the customer.
13:22 So who drive this?
13:26 And then number four is about the shipping. Are we delivering?
13:30 Any delay?
13:32 And the last one will be the what-if scenario.
13:34 Um, we will cover more detail about this. So this will be kind of the same screen we already cover.
13:42 For the CEO, it's the two sentences for CEO. The headline, the best and, uh, the worst.
13:50 So from the narrative, you can see that, oh, the business generate that much. Um, that is the profit.
13:57 That is the gross margin, twelve point percent.
14:01 Um, and then let's talk about, um, the other mode. So, um, uh, anomaly and root cause analysis. If you look at the, the, the product, product page. Uh, let me open the live here. Give me one second.
14:18 So let's say you have a, a chart here. It has a heat map, um, across all the year and different month by the product here.
14:28 You select the anomaly root cause, generate. So hard work at the back end. Uh, I- I will cover in more detail later. But let's look at all the key anomalies found here.
14:40 So it, it say that on the ninth of March twenty twenty-five, the profit jumped to eleven K on twenty-three point four K.
14:48 Uh, there are a few interesting point here.
14:51 Uh, look at September twenty twenty-two.
14:54 September twenty twenty-two here.
14:57 Uh, September.
14:59 Wow. They sell it thirty-one thousand, but the profit's only eleven.
15:03 Uh, something we need to pay extension to.
15:06 Uh, blah, blah. Yeah.
15:09 A lot of detail there.
15:10 Um, so if you… So basically, they're different mode for different audience. Executive, you will have a summary. Anomaly, it will go deep dive into what happening in the back end.
15:23 Uh, cool. Uh, let me move to the next one. Next one about the customer analysis. Uh, let me open the, the, the live version. Customer analysis. Yeah. So here, wow, we've got the chart here for different region, and then we got a scatter plot here for different customer and the customer ranking as well. And then, uh, let's say, oh, I want to find out more detail, any outlier.
15:47 Just select anomaly root cause, click Generate, and select the model.
15:52 And then let's see the output from the, the insight.
15:56 Why it failed. Cindy is the one. Where, where is Cindy? I think one of the, the…
16:03 Yeah, here. See.
16:04 Interesting.
16:06 For Cindy, there were a lot of six point six K on a five point six nine zero sale.
16:15 It's kind of the minus hundred percent margin.
16:19 Uh, let's see what's Eric pick up. Uh, similar. He had the higher sale, twenty-five K.
16:24 I think, uh, the top person here.
16:27 But it generate almost, uh, two K loss at a negative seven point nine margin. So something you need to pay extension to as well.
16:38 Um, yeah, we can then go through, like, uh, uh, a few, the insight from, from this.
16:45 Um, so basically this one we have more detail.
16:48 Uh, let, uh, let me move back to the, um, uh, shipping, uh, commission. It just the same story.
16:57 Uh, so basically, uh, you can explore… Oh, this one is the headline narrator. So this is a new version. I just, I just, um, I just deploy and show you later on how it work.
17:07 Um, cool. So, so basically, uh, just now I just cover, like, briefly, uh, how it work.
17:15 Uh, after the AI extension that we're going to cover today.
17:19 Uh, so let's, let's, let's go back to, um, uh, this, uh, this big question.
17:27 So where does the gen AI extension fit into kind of the current, like, AI ecosystem?
17:36 Um, you can see from here.
17:38 I show you here how I- I classify into, like, four layer.
17:42 So at the bottom layer, that is the, the context layer. So that is the Tableau knowledge, the semantic and the metric, and the permission, and then the row-level, uh, security.
17:55 And then on top of that, um, we have the, the view layer.
18:00 So the view layer is where we build the thing with Tableau Studio, with Desktop and with the web and with the Data App.
18:07 And deliver Delivery, how it deliver the insight with the agent in the dashboard where you can ask a question, uh, generate a summary, uh, uh, with the powers of proactive, where it deliver the insight directly into, uh, the Slack, along with the team, uh, by the MCP.
18:27 And, uh, how about our extension?
18:31 So our extension is just the last step. It is the last step based on what you already built. So let's say you already have… You already built, like, ten hundred dashboard. You already public that. It's already available in there. So the extension is just, it's just called the, the last mile narrator. So it build on top of what you already have.
18:55 It read exactly what you are seeing, and it could write a story about that using the model that you choose. So you don't have to rebuild your dashboard, no migration happening. Yeah, that's it.
19:09 We will cover more about the Tableau MCP, uh, uh, in, in, in the next session in, uh, November. Yeah. So, uh, so I think, I think, yeah, there are a few way that we can put an LLM, uh, next to the, the Tableau data.
19:25 So we have a Tableau Agent in dashboard. Uh, anybody here have been exploring the, the Tableau Agent, uh…
19:35 Andy? Chris?
19:39 Cool. Um, so if, if you give a Tableau Agent, so basically it, it's a native panel in, in the dashboard. It managed by Server and it got the, the trust layer built beyond the extended trust layer. It maintain the ro- role-level security and everything here and there. So it makes sure that, um, for example, a manager has set to the West cannot, uh, ask the question about the sale, uh, for the, for the East region, for example.
20:07 So they are, um, built and managed by, by Server. Um, and to the right is the Tableau, uh, MCP and LLM client. So this one, uh, all the application like, um, ChatGPT, Slack, so you integrate with the, the MCP.
20:26 Um,
20:29 yeah, you can config that and connect to that, and you can do all ad hoc analysis, uh, outside of Tableau.
20:35 Uh, but, uh, to make it, um, a fair, so if the agent in Tableau, the Q&A, it already answer your question, I think that's fine. You don't have to, like, create your own, um, you know, uh, extension for this purpose. So you don't have to, like, maintain the source code. You don't have to build and host that.
20:56 Um, if those, in the middle layer, that is the generat- extension.
21:02 So basically it's just an iframe i- in the dashboard.
21:06 For the model, you can bring your own model. It can be Azure OpenAI, or OpenAI's Gemini, um, can be Opus, whatever. Uh, so basically… Or even just the local.
21:18 Uh, if you run any, like, open, uh, model local on your laptop, even you can connect to that as well.
21:24 Yeah, so, so there are just a, a few, uh, idea about, um, how you can put LLM next to, uh, the Tableau data.
21:35 Yeah, so this is just one of the, a few, a few things that, um, we can explore later.
21:41 Like, about because we have the ability to bring in different model like the Luna, GPT-5 with Luna, and GPT-5 with data, so on, for example.
21:55 But then if you explore it, the output from different model, it can be different.
22:01 But does it tell the same story or not?
22:04 Uh, does it answer the right question?
22:07 Yeah, we will, uh, explore that, um, in the next few slide.
22:11 Yeah, so here, I think one of my, my, my favorite thing, uh, about, uh, what the narrator, uh, pick up, uh, during demo. Uh, just now we already mentioned like, um, oh, um, the number one customer lost money.
22:25 That is, uh, similar. So he the top, um, salesperson, but actually he a loss.
22:31 And so the hidden story up here as well, if we look at the, the, the delivery, 46% ship early, but then are more, more than a quarter are late.
22:42 Uh, and a few hidden pattern that is cover as well, uh, like for technology, it show 31K in September, but then the profit only $11. I think that is a little bit crazy.
22:54 Um, cool. Um, so, uh, this here, um, one dashboard, several narrator. Because this is one of the question that some of our, uh, my friend already asked about it as well. Um, can we bring in, like, different narrator into the same dashboard?
23:15 Yeah, that is possible. You can bring in, uh, different narrator for different purpose.
23:24 You can select, like, whether you want to analyze for the whole dashboard or you want to zoom in a specific data point.
23:33 Yeah. For example, you can select even only the customer and ask the, uh, the, the inside, "Hey, give me a summary of that only for the, uh, uh, consumer." So here is how it, uh, pick up, like, Patrick, Martin, John, Patrick, um, for the selected. Basically, it's kind of the drill down. You zoom out, you view the overview. And zoom in, you can select a particular, uh, data point, uh, and then it zoom in into that.
24:01 Uh, two more same fact. Basically, we can bring in, like, two narrator side by side with different model. Even it answers, um, the same story, but I, I will show you how it answer in different way, even both answer are correct.
24:19 Uh, then we will explore a few scenario, uh, in, in, in, in a demo, uh, uh, later. So here another thing. Um, we have one dashboard and, uh, I let, I think let me, let me open the live here, so it's easier. So that is the new version. Yeah, let me open the new version here.
24:39 Okay, let's say here your dashboard, you want to have one, uh, narrator here on the left and maybe another narrator here on the right, for example, only for… Oh.
24:51 Uh,
24:54 uh, I think let me, let me do it now.
24:56 Uh, Tableau extension. Let me do the local here.
25:00 Yeah. So this is how, uh, you, uh, bringing your extension into the sitting Tab- Tableau dashboard. I'll show you how to work later.
25:10 Um, let's please open the extension that you already developed locally, and how you're going to configure, uh, the setting, uh, for the extension.
25:22 Uh, so the icon setting, uh, inside, you have a full control of that.
25:28 Whatever setting you want here, you can configure that. Here an example, like, oh, um, what do you want to say? You want what it is planner or you want a second opinion?
25:39 Uh, let's say if you said it's a headline, uh, you want to see in the compact format or in the panel, would be the full narrative view.
25:46 Uh, what be the inside more. Um, you want to see the executive summary anomaly. So at the back how it work, we just have like different, um, system prompt for that. If it's executive summary in the prompt we send to the LLM, "Hey, give me the summary in two bullet point." If it anomaly, then we give more details. So you can… You have the full control, um, to customize it, uh, in your own way.
26:12 Uh, you can select like, oh, either the entire dashboard or the total sale, total map or a particular worksheet. You can select the model you want.
26:21 Uh, here I just config only for three model. The other ones they haven't done yet. Um, and then another thing, how you can refresh.
26:30 It generate the insights what happen next.
26:33 If you want to see only for, uh, the central region user, change the name or change the field. You can even configure here.
26:41 When the filter change, refresh the insight, or when the map selection change, refresh the insight, or you want to do it manually.
26:50 Yeah. So I, I will show you how it work later, so you can build your own, uh, extension, um, like this.
26:57 Yeah.
26:57 Click save configuration.
27:01 So what next?
27:02 When you save this configuration, it will store into a Tableau workbook. So next time when you open that, it will have a setting saved. So you don't have to, uh, configure that again.
27:13 Cool. Uh, while it's generating the headline, I think, uh, let's start with this one. Um, I will cover a little bit about Tableau extension here.
27:22 Um, so if you have any questions, just let me know, uh, or I can, uh, take that at the end of the session.
27:28 Um,
27:30 so we have, uh, the dashboard extension.
27:34 So that is what we are going to cover today. So that will be just the web application in the dashboard, like what you see here.
27:41 Uh, let me show you the, the, the, the, the one I already deploy. I think this one. Yeah.
27:50 So this is the one I already, uh, deploy, uh, to, uh, um, Tableau Cloud.
27:57 So this is the extension, uh, on the left here.
28:00 Uh, yeah. So basically it just kind of the iframe, the web application embedded, um, uh, into the dashboard. And then what does it do?
28:09 It can read everything in the worksheet. Uh, let's say here we have like three, three worksheets, for example. You got the product drill down, and then you got the sale and profit by product name. So it can read everything o- o- on, on the dashboard here. So what you see, it can read that.
28:25 Um, and then it can listen to the event as well. When you click on a filter or when you click on, um, a little point, it can recognize that, and then you can, um, have action accordingly.
28:39 So that is the, the dashboard extension. And then we have the big extension as well. The big extension is you can create a kind of custom map, D3.js or whatever. You can bring in, in the map, uh, card into Tableau Desktop.
28:51 And then we have the, uh, the analytic extension. So that is something with the, like the TabPy, you connect to the Python and R for external service.
29:00 Um, so there are kind of the three different type.
29:04 And, um, another distinction is, um, the, the sandbox, uh, um, extension. So that is hosted by Tableau.
29:13 So they no network, uh, calls out.
29:16 So they accept by default, but you cannot like call, uh, an LLM or something outside of that.
29:23 Uh, for our extension, that is called, like a, network enabled.
29:28 So you host that, it can send the data out, but this need a site admin to approve.
29:36 So basically you, you can configure your, your, uh, Tableau Server here to allow, um, to allow big extension.
29:45 I'll show you how it, uh, deploy, how to configure that later. Yeah.
29:50 So for that, you host that, it can send the data out. You need to approve that in the allowlist. And the viewer who load the website, um, they need to, uh, approve that as well.
30:00 Yeah. So, so those are kind of the basic idea about, uh, the extension.
30:06 Um, so these are the, um, the key part for, for, for, for today.
30:10 Um, uh, so I'm going to cover a little bit about, uh, the, the project, uh, what will be there.
30:18 Um, uh, the structure of that and, and how to deploy that as well.
30:23 Um, so this is an example Um, of the project. I think let me show you from here. I got the– Let it open here.
30:32 Yeah. I- I'll show you how to build this in, in five minutes, um, with an AI agent. It's just a simple extension. Uh, it can click and read whatever you click on.
30:43 Uh, so that is this one.
30:44 Um, no more selection.
30:48 Yeah.
30:50 Uh, I will show you how to do it in five minutes with, uh, with, with, with, uh, an AI, AI agent coding. Uh, just to show you the concept and how it work.
30:59 Um, so, so what are the… Uh, let me go back a bit. So what is the structure of the application? Um, so for this project, uh, this is how, how it's just my experience, um, uh, how I structure this thing.
31:13 I got, um…
31:16 You can see that we have, like, uh, the TS file. Uh, it's kind of the Amazon kind of, um… So basically, it's just the manifest file.
31:24 Uh, I have two file here. One is, uh, population, the local one for local development. So this is the file that, uh, I just click on to import into, um, the dashboard just now.
31:38 Click on extension.
31:41 You can try and drop it anywhere.
31:44 Uh, click here.
31:46 Mark Inspector. So that is the extension file. If you click on that, it will, uh, load the extension.
31:54 Cool. So that is, um, if– On the right, you will see the detail of that, how it look like. Basically, it just kind of the pointer. It will tell you what is the ID of that extension, what the name of class, uh, where the location. I think this is important. I'll show you how you can deploy that, uh, a little.
32:13 And then, uh, the permission.
32:16 Uh, here I deploy it like full data so that the, the, the, uh…
32:21 Oh, not this one. Uh, so your extension here, it can reach the whatever you have, uh, on the dashboard here.
32:31 Uh, what next? Uh, we have the extension UI.
32:35 So extension UI, uh, let me show you. Uh, yeah, maybe we'll go back to that later. So we got the extension UI. So that is kind of the UI, uh, component where you config the pop-up just now I show you.
32:50 Um, and then you– where– that is where you config, um, how to call the back-end as well, the link to the back end.
32:57 And importantly here, I got the back end. So that is the, um, container app.
33:03 So that is where I, uh, process all the AI query, generate the insight. Uh, I generate the system prompt, uh, process the, the input and the output. And, uh, there are a few things there as well that I will, I will discuss later.
33:20 Uh, for example, like telemetry, where you can, you can log, let's say, uh, um, what model being used, the token being used as well, and estimated cost for that.
33:31 Um, there are a few things about security as well, uh, how you can prevent kind of the, uh, uh, prompt injection, uh, from the, uh, the, the output it return.
33:43 Uh, yeah, I just got a question. Uh, so, um, the extension, the Docker file, the Docker file for the back end, uh, for the back end. So this part. Yeah.
33:55 So that is how I deploy, uh, the, the back end for the extension. Uh, so the Doc- Docker is only for that.
34:03 Uh, for the front end, uh, it does not require that. You can just, uh, uh, deploy directly into, um, kind of a static, uh, page. Yeah.
34:14 Uh, here are just a few other thing about, uh, benchmarking as well.
34:18 Uh, let's say… Oh, uh, I will, I will, I will cover a little bit about that because here you can see that we have like, uh, uh, different model here. Each model, it will generate different output. How you can do like a quick benchmark, uh, class. Again, the data you already know. Uh, I will show you just a quick example how to do that as well. Uh,
34:40 yeah, so I think that… Oh, yeah, and then we got the, the workflow here.
34:44 The workflow here is how, uh, we do the, the CI/CD. So I got one for deploy the UI.
34:51 So what it mean is every time I make a change to the, to the UI extension part here, um, and then I want to deploy that, um, I can commit to GitHub.
35:03 Uh, I can create a branch with a certain task, and that will automatically publish to, uh, the place where I host the extension.
35:11 Uh, the same for the Docker. The Docker, I use it deploy, uh, API.
35:16 So you can, you can do that manually, um, uh, using any, uh, container service from Azure or Google or whatever.
35:25 Um, you can do manually, or you can use the same way, uh, with the GitHub Action.
35:31 Um, yeah, and then a few rules for the agent and security, what you should do, what you should know. Yeah, for that.
35:38 Uh, yeah, so those are the, uh, basic thing, uh, about the structure of the project.
35:44 Um, next, maybe I will cover a little bit about how, how it work.
35:50 Um, so here are just the, the five step.
35:54 Uh, first, just now, uh, I show you how it work. Just now we just, uh… Ooh, wait. Yeah. For example, if you see one atten- extension you just create. The first step is you click on extension, drop it here, select a file.
36:08 So it will, uh, open the URL that in an iFrame.
36:14 That's just the first step.
36:16 And, uh, the second step, you need to co- initialize agent.
36:20 So it kind of the handset, uh, with the host.
36:23 So that would be, uh, the third thing.
36:27 And that need to be call, um, before the other step.
36:32 The next step will be, uh, the discover. So why discover? So basically, it just, um, get the list of all the worksheet on the dashboard.
36:41 Uh, here, here just a simple simplified version. You can see that, oh, that give me all the worksheet on that dashboard.
36:51 For the example I show you here, it will just return all the worksheet, um, on the dashboard.
36:56 Like we got like a few here.
36:59 Uh, so next, read the data.
37:03 Uh, you can call the function to get summary data as in. So with this, the filter already apply and the row-level security applied well.
37:11 Um, I think this is, uh, an important thing to, to, to remember. So the data it get are what you supposed to see and what you see as well. It does not get the data from the other user or from the other region that, um, it not seeing.
37:28 Uh, and then I think the last one will be it will trigger after it get all the information, it will generate, it will call the, the LLM, uh, on the server side, uh, to generate the insight. Uh, I will show you a quick demo later.
37:44 Um, so that is, uh, this slide.
37:47 Uh, after you get the insight from the server, and then we just render that in the inner HTML. So here there are a few things to note about security and about how to optimize the token, uh, user.
38:03 Uh, so here you see that here I return the JSON.
38:06 But instead of JSON, for JSON for every row, if you look at the JSON file, if you repeat the column name, uh, property name again, so it will, uh, consume more token compared to the CSV. So to optimize that, uh, you can convert serialized addition into the CSV instead.
38:27 And also a few security we have. Well, I'll cover that later, uh, on how you can sanitize the input before– as output before you, you render.
38:38 Yeah, so this is, uh, I think I already covered this.
38:41 Um, so you have…
38:44 There is something called setting, uh, API.
38:46 Setting API. So how, how does that work?
38:49 How does that work?
38:51 Um, let's say here one example. You got one, um, narrator here at the top and one here as well.
39:01 This is a setting you already configured. You close the workbook, you open, it still load the same setting again, uh, for you. So you can have one setting for executive, uh, one setting here, um, for like big data analysis, for example. And you can have another one here, for example, um, with different insight mode.
39:22 Uh, so you can check it out, um, the setting, uh, API. So that is, um, how you can configure the different setting and push it back into the workbook.
39:34 Um, yeah. But, but one thing to note is don't store any, like, secrets here, uh, in, in, in, in this setting. Like, uh, just, uh, have some key here, API here, whatever here. Just don't, don't store that here. Yeah.
39:52 Uh, cool. So, uh, let me, uh, go to the next, uh, next slide.
40:00 Cool. So the first thing about the summary.
40:03 So summary, not underlying. Summary mean, uh, when you click the generate here on this, um, or this overview, for example. So whatever you see here in the summary, it will take and then it will send to the LLM. So instead of sending everything at underlying data, so that will take or consume a lot more, uh, token.
40:26 And then another thing is you can choose to send the column name as well.
40:31 Uh, for example, in the, in the older version, I didn't send the column name, I just sent the value. So it tried to guess, oh, the column appeared to be like profit margin, profit and sale.
40:42 Um, but you have the option, uh, to send both the field name and the field, uh, value as well.
40:49 So the, the LLM, it, it doesn't guess, oh, it is the profit or it is the, the sale.
40:55 The last thing you react to the user. So basically, you have a lot of option on how, uh, you can, uh, refresh or react to the user when they interact with the dashboard. So when they click on the mark, what will happen?
41:10 Uh, when they click on filter, what will happen? So you can refresh the, um, the insight again, or you can have a delay for a few second before you refresh.
41:21 So that's, uh, up to you.
41:23 Yeah, another thingy about the token, uh, I already covered here.
41:28 Uh, so it's better to, uh, use CSV instead. So that will save you a lot of, uh, token instead of using the JSON.
41:36 Because the JSON, it will, uh, repeat the property name for every, uh, record.
41:43 Cool. I think with that, I'm gonna show you a quick, uh, example on how you can build, uh, an, a extension here, um, with, um, the AI coding agent. Um, I think most of us here already an expert in, in, in Claude, um, and how it work.
42:02 Um, for me, I also explore the Google Antigravity as well. Um, so, uh, I will show you a quick demo, uh, on the Google Antigravity. I think it also have the, the AI, uh, coding agent where you can use to, to develop all this one. Um, so basically it's two different, uh, project, but both work well. Uh, so I think the thing that matter will be the, the, the, the interest and the skill that you give to, uh, uh, the AI agent Uh, so on, on the cloud, uh, it could be like, uh, the CLAUDE.md.
42:36 Uh, so it key every session the project group. So for the Google Earth activity, it also have a similar thing.
42:42 Um, it c- it can run a several agent in parallel as well. It can ta- list the task, plan. Um, it also have a browser agent, uh, as well, um, uh, for testing the UI.
42:54 Uh, cool. So this will be the, um… I, I, I will share this one later.
42:59 Um, so this will be the file that, that we can, we can try. That will be kind of the rule that we give to, um, the cloud or the Antigravity, uh, how it work. Uh, so you need to load the extension from the library because there are many time like, oh, okay, create the extension. It loads the library from somewhere else. It's not the library person, for example.
43:20 Uh, you need to call it before each one. So basically, the, the, uh, all the rule, uh, that you, you will define, um, uh, for the agent. So for, for the example demo, um, I add the agent, uh, to create this, uh, extension, the extension. You get the local one. It cannot connect to the internet.
43:38 Um, so what it doing, uh, if you click on, uh, any mark here, any mark here, we, we call it the mark inspector. Yeah. You click on any mark here. Uh, if you recognize that, if you say like, "Oh, only one mark selected," it have the lat and the long, and what the value of that. Uh, you can select like multiple as well. Uh, one second.
44:00 You can select the multiple mark here as well, or you can recognize that, and they just display, uh, in kind of a table format. Yeah. So basically, just a simple, uh, extension.
44:11 Uh, you can reach whatever data point or mark you select here and then visualize there.
44:17 And then I show you how to create this, uh, with just, uh, uh, one prompt, uh, one rule.
44:22 Um, it may not work the first time. I try, uh, on the Google Antigravity. I need to, like, tweak a little bit.
44:28 Um, but let me show you a quick, uh, quick, uh, uh, a quick video how it work. Mm.
44:37 Yeah, so that, that, that is the interface. Uh, so that will be the rule. So we add it to add the rule to the project.
44:43 Uh, and then it will initialize the project for me.
44:47 Uh, that is the rule, blah, blah. And then next step, I add, "Hey, let's build the dashboard extension called Mark Inspector," blah, blah, something.
44:55 Uh, then within a few minute, uh,
45:02 it, uh, create the extension for me.
45:05 Uh, and then that will be the readme. I add, "Hey, run the local server. Add the extension to Tableau Desktop, and then interact with the marks." Yeah, that is what we did just now.
45:12 Um, ooh.
45:17 Yeah.
45:18 Uh, so that is what it generate, uh, before. Uh, let, let me show you what it, uh, generate, uh, without a prompt.
45:28 Um, yeah. To me, what, um, the, the, the agent generate, uh, so what do we have? So we got that is the rule.
45:36 Uh, we got the mark inspector, the threat file. So that is kind of just the, the pointer we have. It has the extension ID here, the name here. And I think it important, this is where, um, it's kind of… For the demo, I will host this extension locally on the port 4000.
45:59 Yeah. Because when you add the extension into Tableau, uh, Tableau will try to make a call to that local host 4000 index.
46:09 Um, yeah, I think let me… Yeah.
46:13 There was l- uh, yeah. And the index.HTML, so what is it?
46:18 Yeah, so that is the interface, uh, for this one. Basically, just the HTML page. Uh, you can decide what you want to display, in what color, what format, whatever way you want.
46:29 Yeah. So it look very, uh, more straightforward.
46:33 Uh, and how to serve it, um, the extension. So I just start the server here.
46:38 You can just call it Python HTTP server 4000 because I already have the other one on 3000. That's why I just use 4000 here, just locally, uh, for testing.
46:48 Um, yeah. Uh, you can see that, uh, every time you click here, yeah, it will update the table, uh, accordingly.
46:59 Yeah.
47:00 Yeah, you know, you can see the update here.
47:04 Yeah. So that, that basically just, uh, um, a quick, uh, uh, demo, uh, for this on how you can build the, the, uh, extension here with just a single prompt.
47:17 Uh, I think let me, uh…
47:20 Yeah.
47:21 Yeah. So for that extension, um, this one, it's just a simple one. It did not connect to the internet. It did not, like, call any LLM.
47:31 Uh, so for the AI extension, the one that, uh, uh, we hit just now, here will be a more detail of how, how it work.
47:40 Uh, on the left side, you can see when you click on, um…
47:44 Yeah, so on the number one, uh, in the dashboard, when you click on Generate, uh, what happen?
47:51 Uh, if you look at the number two, the number two is the, the one I deploy on, uh, Docker, Kris.
47:57 Yeah. Uh, you… Uh, I think, let me see if they have that, uh, deployment here.
48:02 Uh, or is it AI? Yeah, uh, maybe, uh, let me show you over here. Um, so, uh, on the right is the, the, the back end that I will deploy, uh, to Azure, um, uh, using the, the Docker.
48:18 Uh, you can do the same on a Google or, or AWS.
48:21 So every time you click on that, uh, what will happen?
48:24 Basically, it will make a call, uh, to the API here.
48:30 It say, "Hey, uh, generate the insight for me. Here are the data." The deployment, uh, the, the, the, the model selected the GPT 5%. That will be the endpoint. I already deployed that. Yeah. Um, I, I show you the local version. Uh, number three, just type the local version for testing.
48:48 Uh, if you deploy that locally, uh, you will see something like that.
48:53 Yeah, I think this is the, the local version of the, the, the, the API that I deploy, uh, locally. So the same thing, um, I deploy, uh, on the, the, the Azure. So when you click on, um…
49:09 Uh, give me one second. Let me say, yeah.
49:13 So this is the, the local version, um, for the API that I deploy, uh, locally here. You can see that, uh, it, it, it's serving, uh, request here. Wait, idea, idea. Wait, wait, wait. One more.
49:28 Yeah. And now,
49:32 uh, so for testing, uh, I export like three, uh, method here, health check, what happened in there, and either HTTP post generate insight. So that is where it will accept the request, um, uh, from the extension. Hey, generate the insight for me. That is the model. Uh, that is the, the input.
49:54 That is the API itself. Uh, I will, I will, um, create a sample for this, and then I will share it with, uh, our group, uh, later on, on how it work. Yeah.
50:07 Um, so I think here just a few thing about like, um, the lesson learned, uh, for me. Um, so I think, um, because when you deploy to production, so it will be the HTTPS endpoint. Locally for testing, uh, it's the HTTP.
50:25 Uh, so, so sometimes for testing the HTTPS cannot call the HTTP, uh, so it's ca- it cause some mixed content, um, uh, error there.
50:34 And a few thing with the K8s reload as well. Uh, and I think this is important, the secrets key.
50:42 So we don't store the secret key on any front end. I think sometimes for some agent, it just stores the key somewhere, and then just put to, to production.
50:51 I think that is a, a very, like security risk.
50:54 Um, so for, for security, for key, we either, either using keep on the server side only.
51:02 Uh, so there are a few, um, like the, the key management on the cloud provider that you can, you can leverage.
51:11 This is another thing, like it's, it's the CORS policy. So basically just block the origin.
51:17 And hey, this is not from the same URL, just block it.
51:22 Um, so we did a few thing when you generate the extension, uh, it may fail the first time, um, for certain reason. Uh, so to, to debug that, uh, you can use the… You can learn, uh, in the desktop with the debug mode there to see what are happening, uh, in between.
51:39 Um, here a little bit about the back end.
51:42 Um, or how, what's exactly the, the, the container that, uh, I think Chris asking. Um, uh, so, uh, you got the extension UI, so that would be, uh, uh, the left.
51:53 When you click on that, what happen?
51:56 It will just trigger an API, um, to the server.
51:59 You can see the server here. So in the server here, I explore like a few methods, like the health, model, or generate the insight. Uh, I think an example just now is, uh, yeah, we got like three here. Um, so that is the local one.
52:12 Yeah.
52:13 Uh, the model just return or the model available that I already configured. Um, the health will just do a quick check if everything okay. Uh, generate insight will be the key thing.
52:23 Um, when you generate the insight, uh, you need to provide what model, uh, you selected. Like that, either the Soluna, Mini, or whatever. So when you click select on the extension, it will just pass that as a parameter.
52:41 Um, and then you need to complete the, the, the key as well.
52:44 Um, so I think, I think, uh, one thing here when, for development purpose, um, the key you may store on the environment, um, available. But for deployment to the Docker, um, so you can better configure it over there, uh, instead of just store, um, in the flat file.
53:07 Cool. So I think this is the, uh, thing we talk about how we configure like different mod- different insight mod, uh, for the agent.
53:17 Um, you can, you can see from, from, uh, the list here, like we got like different, uh, uh, insight mode here. But, but in the back end, how does this work?
53:28 Uh, so basically it just kind of, um, here how you define the prompt template.
53:34 So if it's the executive summary, you add the LLM, "Hey, uh, I just want a two-sentence, uh, executive summary." Uh, that's it.
53:44 Um, um, uh, if for the anomaly, "Hey, you just find out what are the outlier and the likely root cause for that." For recommendation, uh, you can provide like two or three actionable moves, something like that. So basically you have the full control on how you can configure, uh, this. You can define whatever skill or model that you want, uh, and then you just send that to the, the, the LLM.
54:10 So that is, uh, how we can, uh, structure, uh, the prompt template before you send to the LLM.
54:17 Yeah, so here just a, a few thing, um, um, when we…
54:22 So from my experience when I test out, uh, so always you ask for the structured output, because sometime you call the LLM, it might return the data in different format. Uh, and then on the client side, we cannot like parse the data because it have no idea, um, where to get that number.
54:39 Uh, so you need to Our first structure are both that you can pass it. One is the return.
54:46 Uh, business re- business definition. How does it know like, oh, uh, the either revenue that either profit. So this is something like similar to the, the, the, the workbook, the business reference.
54:58 So that the, the LLM it doesn't guess.
55:01 For example, in the older version, like I even I didn't provide the, the column name. So if you try to guess, oh, this would be the profit, that would be the revenue.
55:11 Uh, so you can provide that in, in, in the description or some, some business rule.
55:17 And another thing about the, the, the– for different model.
55:22 Because different model, the both of them different. Uh, I will show you how when we run the benchmark, you will see the different, um, in, uh, kind of, uh, for example, in the timing, for example.
55:33 Uh, so for, for like, uh, the mini version, for example, it very quick and fast. Maybe it's good for summary, so you can use that for the summary.
55:43 For now, that require the reasoning, for example, then you can use some, uh, S5 model, but it will take more time.
55:52 Cool. So I, I will show you how you can do this, uh, benchmarking, uh, using a few, uh, available tool. Um, yeah. So if it be more, uh, I will show you a quick example of how you can use the prompt tool. So the either open source CI, CI-LM, uh, to do a quick kind of benchmarking across like all the three model, uh, later.
56:15 Yeah.
56:16 Yeah, yeah. Uh, I think let me check if we have the, the example here.
56:21 And for example, now I already complete like three, uh, three, uh, uh, model here, uh, like Fireboyz, it's John, and Mini, and Luna, and how I'm going to like do a quick kind of benchmarking, um, of the result of the three. Uh, so here just have a simple way, uh, for you to test. Um, I already can run it now. Uh, I will send the instruction maybe later after the, the session.
56:48 Uh, so you can, you can try it out on your own as well. Um, here I just define, uh, a few kind of, uh, the crowd, uh, rule, uh, for, for benchmarking.
56:59 Uh, where the feature, for example, to be in the kind of the crowd rule, uh, based on this. So that are the key.
57:05 The what is the expected number, uh, that is the crowd, uh, fact, uh, by other. Just kind of the simple one.
57:13 Uh, and then, uh, we just, uh, execute this. If you run around like fitting test at that time.
57:21 Uh, and then you can see the result here with the pass or not. I also could fit the, the threshold timing as well. So if it below the certain timing, if you pass or if it, if it fail, um, and then you can just, uh, launch that in, in, in the browser. Then you can see the, uh, uh, let me, let me open that again.
57:41 Uh, where the things is here. Yeah. So here is an example of the, the result of the benchmark, um, of the three different model.
57:53 Uh, so you can see here for the scenario one. So what is the regional profit, uh, anomaly? Earlier record analysis. Um, so what is the worksheet name?
58:02 Uh, what, what is the month that it, it, it true for the benchmark? Uh, what is the role? Uh, it rather, uh, see feed data. So that is the return from the, the CHV.
58:12 So, so what do you see from here?
58:14 Uh, we can see that, oh, uh, uh, GPT, uh, mini, uh, request file. Uh, what is the average latency? Thirty-four second. Thirty-eight second for the Luna.
58:28 Uh, and around thirty-four second for the song. I think this one may be the, the fastest one. And then if you do the quick check of the data as well, uh, and the filing as well, because even the result are correct for three different model, but the output can be different. Because this model might pick up like, uh, this one is, uh, good salesperson but have a loss.
58:51 But the other model might pick up the other case.
58:53 Um, yeah. So that is just one simple way on how we can like do the benchmark of, of, of, of the model. And it should verify again the number that we already verify it at kind of the, the, uh, kind of the setup tool. Uh, yeah, almost there. We have like a few thing to cover. Yeah. So, um, that is how it, how it look.
59:16 We got five scenario, three deployment for, for three model. Uh, mini, uh, Fireboyz, uh, Luna, and then Song. And we got like fifteen test case.
59:27 And that will be the result.
59:30 So this is the one I already mentioned. A path, but it's not the ranking. So if you look at the scenario one, for example.
59:37 So the mini, it will open with California profits. Uh, but for the Luna and the S- Song, it open with Texas and, uh, the, the Tennessee. Yeah.
59:49 So basically they have a different output from different model, even the answer is still like correct.
59:55 Cool. So this is a bit about the CI/CD.
59:58 Um, so I already cover like for the GitHub, I have like two project for the front end. Uh, so I set up kind of deploy UI, the YAML here. And then every time you make a change to, to, to that, uh, and then you push the change to the re- release version, it will automatically, uh, deploy as a static web application. That is for the front end. Uh, for the back end, it will be here.
1:00:18 Uh, uh, so for this one, I set up kind of the pipeline as well. Or you can do it manually, that's fine. Uh, you can build a Docker, and then you can just, uh, deploy, um, using any container service. Yeah, so that is the front end, how it work. That is the, the structure of the file, um, where you can config in, in, in your project. So every time you just check in GitHub, it will just automatically deploy, uh, for you.
1:00:43 Uh, same it for the back end.
1:00:45 Um, on how you can config for the, the CI/CD as well.
1:00:49 Yeah, I think number five, the key, um, where you should keep the secret in the Key Vault or, or the managed, uh, identity.
1:00:58 Yeah, here just a little bit, some different ways about how you can secure your Tableau extension.
1:01:03 Um, so for, for the Tableau side, it already give you almost everything here for the every single sign-on. Uh, and then the safest for the extension, you must allow that. Uh, row-level security already apply so that extension only see, uh, what the user is supposed to see.
1:01:19 On the right side, that would be your job. Uh, it's a lot of work here, right? The first thing you need to protect your API.
1:01:27 Um, by you can fix, like, uh, after you deploy, you must allow the API only accessible from, like, certain IP or certain website or certain origin here. Just to lock that to the extension origin.
1:01:41 Uh, a few other way as well, you should, like, cap the span as well. It can be in exploit, um, there it should reach the limit.
1:01:51 Another thing is you can sanitize the AI output as well before you render in that.
1:01:56 So basically just for, to prevent kind of all the prompt, uh, injection here and there. Uh, you may heard a lot about that.
1:02:03 Um, key on the server side. I think that is also a key thing because there are many agents, they just, "Oh, let's query this one, use this key," publish everything, even on the, uh, front-end side. Uh, so for, for some project, they publish that, and then the user can see what are the key already published, uh, on the server.
1:02:22 Um, generate error message, so you don't log any, like, key, any, um, a bug or any, like, URL, whatever in, in, in, in the error message or PII or whatever. So just don't keep that in, in the error message. So that may be, uh, display in the log, um, on the server.
1:02:43 Uh, so I think, I think, um… Yeah, so I think, uh… Yeah, so here just a few here, uh, for you just to, um, uh, for your protection. Um, so we have like spend cap. And if you can copy that as well on the server side, um, on, on the right, just to prevent the case when it being like abused or…
1:03:05 Secret key, the data as well, uh, extension here.
1:03:09 Uh, so here just a few config. You can check it out, CSP frame ancestor. So it only log inside the Tableau. The doc purify, uh, we already cover that for the injection.
1:03:20 Um, on the API here, you can config the, the origin here and then the limit as well.
1:03:26 Just in case someone find out the key, "Oh, we got the key to this, uh, extension or API. Uh, so let's use that.
1:03:33 Let me send a thousand of request and just use up all your, uh, quota for that." So I think, I think that that's almost, uh, everything about the extension. Uh, so I think I will cover a little bit about, like, uh, the token, how it work.
1:03:47 Uh, so every time you click on the button here, what happen at the back end, and how about this go.
1:03:54 So it's just, um, uh, simple like, um, how, how it work here.
1:04:00 Uh, so the, the image in like just one token will be like, uh, around four character, four character. That's it. And remember, the rate limits here, it just counted in token per minute.
1:04:12 Because when you deploy it to Azure or Vertex AI or whatever, so they have the, the rate limits.
1:04:18 So you cannot send like too much data, uh, within the minute.
1:04:22 So that's the first thing.
1:04:24 And what will be counted as a token?
1:04:27 So the token, it will in… Have like several part. The first is the input that you send. When you send the input, whatever in your, uh, the row, the inside model, whatever you select there, so that will be included in the input.
1:04:42 And then it will have the response data as well. Whatever the data you send over, that also in the, in, in, in the output, uh, in the input.
1:04:49 Um, and then you got the output.
1:04:52 And, and one thing to remember here, the token output, usually the cost is more expensive than, um, the cost for the input token.
1:05:04 Uh, here I just show you one example on, um, the one I extract from the, the Azure on how it calculate. Like I send one request, it had like four thousand three hundred, uh, twenty-one token. So the cost per token, per one million token is around like four dollar. So if you click generate, um, it sends some request, have like four thousand three hundred token.
1:05:27 So if you divide by like… So one token will be like, about like a quarter, three and a quarter in this work.
1:05:34 Um, then the cost will be around like zero point zero one seven three. So that is the input token. But the output token, the price different. So it's around like five time, like twenty dollar per million token.
1:05:46 So if you send that output, input, you get that output all together, that will be your total token. And the total cost will be, uh, around, uh, this at the left, input cost and then, uh, together with the output cost.
1:06:01 So imagine like for one request, it cost about that much.
1:06:05 And if you got like hundred thousand of that, it will cost around like two thousand, uh, for a month.
1:06:11 Yeah. So that is just one, one, one, one simple way on how we can calculate and then, uh, monitor that. Uh, this is just an example, uh, well, for one week, uh, a simple test, uh, I did. So we send that much of request, one point two two four million token, but ninety-three percent input token.
1:06:30 Um, and then, uh, what it be the total bill.
1:06:34 Yeah, I sample some result from the, the, the Azure, uh, where I monitor, uh, the cost.
1:06:41 Yeah.
1:06:42 Uh, I think let me check. I think that's it. Uh, or more Yeah. So that is just a simple, I think that the most line.
1:06:50 Uh, and then so here are just a few way on how you can, like, um, reduce or optimize the token, like just use CSP duplication. So if we optimize the input token, um, you can cap the ro- the row. Like, okay, let's set maximum, uh, how many record, um, for the, uh, for the row and that work well.
1:07:11 Now, here are just a few other way. Token limit producer, you need to do a lot of sub-configuration there as well. Uh, you need to pick the right cycle of model. For self-summary, that'd be the cheaper one. Uh, for the last listening one, you can go for the, the, uh, the expensive one.
1:07:28 And here are just some checklist or, or recommendation, uh, for your, uh, extension. And just a few things you can check it out, like you said, the column name, the type as well, uh, with the one that we used to pick.
1:07:42 But important thing, the number point, three, uh, is we need to sanitize the AI output as well. You can use a few libr- library. One of that is the DOM Purify just to sanitize the output before you render that. Because sometime we can just set some, uh, adjustment in the prompt as well to execute, um, the judgment.
1:08:03 Uh, and then you can lock the API to the origin to for- so that the other, uh, if they found out about, oh, here is where you host that, but they still cannot execute, um, uh, the API.
1:08:15 Uh, you can set a budget as well, uh, and then you can deploy with the CI/CD. So, so I already showed an example how it work.
1:08:23 Uh, and benchmarking on every pro- or model gen, uh, just to monitor, uh, the quality.
1:08:31 Uh, yeah, so I think, I think that's almost, um, at the end of my, my presentation today. So I think that'll be kind of the three key takeaway here. Um, uh, so first it'd be the HTTP API and Studio. That is for kind of the new work that help you to build the thing. The exten- extension that you build here will be the last layer that build on top of that.
1:08:54 And the number two is a good contact file, it could be the better model. So you can plan for that, you can review, and then you can run that.
1:09:03 And then the, the third point will be about the CI/CD, uh, some tip and trick on how you can do a CI/CD.
1:09:09 Uh, and then some security, uh, on how you can, uh, protect your extension. And the token metric from the day one, so you understand how it monitor and how it work as well.
1:09:24 Uh, here, I, I will share a few, a few resource on, on the website. I, I, I will checking, uh, some sample code here on the website. I will share the link later. Um, I will also have a list of, um, of the resource here, uh, for you to get started.
1:09:39 Um, so there are some, uh, a free Sandbox, uh, program here.
1:09:43 Uh, if you want to, uh, uh, create your own extension, you want to test it somewhere, uh, you can, you can get, uh, registered here. You will get one kind of the free Sandbox site kind of, uh, for testing.
1:09:55 Uh, you can join the Data Dev community as well.
1:09:58 Um, and here be a few link about the next bot extension.
1:10:02 Um, I, I will share this, uh, uh, uh, later.
1:10:07 Here are a few thing for the MCP and then the, the, the, uh, stop, uh, that is for the session in November, December. So, but I, I will share all together. So if you have time, you can, you can, uh, check it out.
1:10:21 Um, yeah. So I think that has come to the end of my, uh, uh, presentation.
1:10:28 Uh, if anyone had any, uh, question, I think I hand it over to, uh, uh, Chris.
1:10:37 Chris McClellan: There was, there was a lot of stuff in there.
1:10:40 David Nguyen: I do.
1:10:42 Chris McClellan: Um, with- without going through each one individually, because I think, um, you know-
1:10:46 David Nguyen: Yeah
1:10:46 Chris McClellan: … my question right at the start was about governance and security of data, and you s- you-
1:10:50 David Nguyen: Yeah
1:10:50 Chris McClellan: … answered some of those along the way. So the, the data-
1:10:53 David Nguyen: Yeah
1:10:53 Chris McClellan: … has to get sent to the model so it can understand-
1:10:57 David Nguyen: Oh
1:10:58 Chris McClellan: … you know, the context and stuff.
1:10:58 David Nguyen: Yep, yep, yep.
1:10:59 Chris McClellan: The question, basically.
1:11:01 David Nguyen: Yeah.
1:11:01 Chris McClellan: Um, which is sort of why a lot of people are worried about what goes out to models and to, you know, self-host this stuff anyway, to begin with.
1:11:10 David Nguyen: Yeah, yeah, yeah. Yeah, right. Um-
1:11:13 Chris McClellan: At one stage you had the, not the mark inspector, there was the other one. Was it the commentator or the one that was doing the commenting?
1:11:21 Um, and Jacob said, "Could you use dynamic zone visibility and user group filtering to detect if the user is an executive or an analyst and show different narrative goals depending on who's looking?" So you know how you were actually choosing the goal and then-
1:11:38 David Nguyen: Oh, oh, yeah
1:11:38 Chris McClellan: … choosing the goal and choosing the role?
1:11:41 David Nguyen: Yeah. Yeah, yeah.
1:11:41 Chris McClellan: Is could you…
1:11:43 I, I think the answer would be it's your extension. You can make it work the way you want, right?
1:11:48 David Nguyen: Yeah, yeah. Possible. It possible. Uh, because from the ex- Actually, let me go back to this. Uh, the, the… Yeah. Um, y- you, you can, you have full control of that.
1:12:00 Um, so from the extension you can detect like who will log in.
1:12:03 And then I think from the I- API side as well, uh, you can find out and say who, who is that user, uh, and then you can just, uh, filter for the right one.
1:12:14 Uh, that, that, that possible. It's simple. Yeah.
1:12:18 Chris McClellan: And while you're on that, that screen, where you're saying with Fast API you've got the, there's an endpoint of health models-
1:12:25 David Nguyen: Oh, oh
1:12:25 Chris McClellan: … generating side.
1:12:27 David Nguyen: Yeah.
1:12:27 Chris McClellan: Did you come up with those names or that's like the standard name that Tableau-
1:12:31 David Nguyen: Oh, oh.
1:12:32 Chris McClellan: Just to-
1:12:32 David Nguyen: Uh, no. It, it, it your own thing. You can do whatever you want. Yeah.
1:12:35 Chris McClellan: Oh, okay.
1:12:36 David Nguyen: Um, yeah, yeah. So, uh, I, I, I will send, um, I will check in a simple, uh, like, um, server side, a back end for that.
1:12:43 Chris McClellan: Yeah.
1:12:43 David Nguyen: Um, uh, so you can do whatever you want. You can have extension here like, oh, uh, let's say you give the thing, uh, okay, since the server side, server side do whatever thing- On, on the server side, they just return. So here is just one, uh, API. I call it the generate insight-
1:13:00 Chris McClellan: Mm-hmm
1:13:00 David Nguyen: … where I have the function to call the LLM, get the insight. But you can have your own thing. You can generate a graph, or you can do whatever you want there, uh, in your own way. Yeah.
1:13:12 So you have full control of that. Do whatever you want-
1:13:14 Chris McClellan: Yeah
1:13:15 David Nguyen: … uh, on the server side. So it open a lot up, like, uh, uh, you know, uh, flexibility, uh, from what you can do, uh, on the server side. Yeah.
1:13:25 Chris McClellan: I know. That's pretty cool.
1:13:27 David Nguyen: Uh, yeah.
1:13:28 Um-
1:13:29 Chris McClellan: Um, Jacob sort of answered his own question.
1:13:30 David Nguyen: Let me see. Let me-
1:13:30 Chris McClellan: I think I did too.
1:13:34 Yeah, that's, so that's all the questions from the chat.
1:13:35 David Nguyen: Yeah.
1:13:36 Chris McClellan: Any, any other questions from the audience at all?
1:13:40 I think I know why you wanted to demo this on a Friday afternoon, right? So we've got the weekend to play with it.
1:13:46 David Nguyen: Yep, yep, yep, yeah.
1:13:48 Uh, I think, um, yeah, I, I think, uh, I… Let me, um, I, I will prepare, like, all, all, all the sample, um, demo, and then I will check in the GitHub as well. Uh, and then I, I will share with everyone.
1:13:59 Chris McClellan: Sure, yeah.
1:14:00 David Nguyen: So that you, you can play around. The, the first API is at the, uh, LLM. Um, there are some free service available as well. Uh, like with the Gemini, you can, you can play around with that. Um, but, uh, if you do not have, uh, if you use, uh, the one deployed by Azure, for example, or the WordDesk Enterprise, so there are some kind of data government, government about that.
1:14:21 Like, they cannot use your data for training.
1:14:24 Um, there be something like, I think, uh, the 30-day, uh, I think, a few, uh, uh, re-re-retention, I think, something like that.
1:14:34 Chris McClellan: Mm-hmm.
1:14:34 David Nguyen: Um, but they cannot just use your data, uh, you know, for, for any training. And when you deploy your API as a your, your, your, your AI here, there are a few options for you to choose from. Uh, for example, here I deployed to the, the region, um, uh, where the Australia, Australia east only. It mean, uh, when you send the data, it will only send to the Australia east region only.
1:15:00 It did not, like, go out of the, uh, uh, Australia country.
1:15:05 Chris McClellan: Yeah.
1:15:05 David Nguyen: If you choose to deploy that, uh, the endpoint, the global one, then your data may go out of the country. Um, that may be a problem, uh, you know, for the, um, uh, like for financial data, for example. They may not allow to do that. Uh, or even here, if you want to do the local, I think that's fine. Um, you can deploy all your local endpoint somewhere.
1:15:25 You can test that locally, so it did not, like, um, send the data to any, uh, LLM. Uh, but if you use enterprise version, there be some agreement, you know, between the, the, the company and, and, and the, and the-
1:15:38 Chris McClellan: Mm
1:15:38 David Nguyen: …
1:15:39 uh, enterprise. Like, oh, you cannot use our data for training, uh, and that, oh, you must guarantee like 99.7, maybe uptime, uh, all those kind of thing. Yeah.
1:15:49 Um…
1:15:50 Chris McClellan: Yeah, but it feels like there's a lot of configuration there just to, to be able to maintain the costs, but also-
1:15:57 David Nguyen: Y- yeah. Yeah, yeah
1:15:57 Chris McClellan: … let, let enough people use it, depending on what your budget is as well.
1:16:03 David Nguyen: I think, I think, um, the, the first thing will be the… Let me, let me go, go back to this here.
1:16:08 Uh, I think the first thing will be the, the governance, uh, like, like what you mentioned, Chris.
1:16:14 Chris McClellan: Mm-hmm.
1:16:14 David Nguyen: Like, a lot of company, they still, like, afraid. Like, uh, even they don't, even they, they afraid to turn on the Tableau AI, because they have no idea, oh, what is behind the Tableau AI. So it's a blocker.
1:16:24 Chris McClellan: Yeah.
1:16:26 David Nguyen: So if you give the extension with e-even the external LLM, oh, what will be the security risk here? So I think that'll be the first thing, um, you need to sort out. And the second thing, about the cost.
1:16:38 Um, the cost, I think there are a few thing, uh, involved here.
1:16:42 Let's say for the extension, for example, uh, we talk about the local container. That will be the cost on how to make, um, maintain that on, on the, uh, uh, Azure container service or whatever there. That is one part. The second one will be the LLM cost. So they need just one example, uh, for the LLM cost. So it, it provide an example on how you can calculate that, so you can estimate, um, how much does it cost.
1:17:08 For example, like the example, like, for, for, for the week, I send that much request, and I, I consume around one, 1 million, uh, token. So it cost around, like, $11 based on this. But if you have, like, 1,000 user and that's more the data-
1:17:23 Chris McClellan: Mm
1:17:23 David Nguyen: … then you can estimate all the cost, uh, how, how much does it cost. Um, the other good thing is you have full control of that. You can bring in your own model. But the thing here about the maintenance, the security, data governance, um, and then, uh, some other thing that you need to, to pay extension to. Uh, but if you want something really simple, uh, Tableau already has a solution for that, like the agent Q&A.
1:17:46 And then you just turn it on, and that's it. Um-
1:17:48 Chris McClellan: Yeah
1:17:48 David Nguyen: … whether you want to have, like, full control, uh, like you want to do something on the back end, uh, uh, to customize that, do it in the way you want, uh, then you can explore, like, this, this kind of a solution. Yeah.
1:18:04 And another thing here about the benchmarking, also an interesting thing, because we explore a different kind of, uh, model, but then there'll be a different way for you to, to benchmark the result.
1:18:14 Um, I, I, I will share how you can do that on your own, uh, and then you can explore a few API, and then how you can be-benchmark the result. Because the thing is, um, you need to make sure that when you deploy something new, it did not, like, break the, you know, the, the current logic. Um, yeah.
1:18:32 Chris McClellan: And, and Paul's comment was he needs a Mastering AI for Tableau Extension for Dummies. Yeah.
1:18:39 It's just-
1:18:39 David Nguyen: Sure. Sure. But, yeah
1:18:40 Chris McClellan: … it, it's the same as a lot of AI stuff. There's, there's a lot of, sort of a lot of assumed knowledge. There's a lot of structure to get going to actually make it work.
1:18:50 David Nguyen: Y- yeah, yeah.
1:18:52 Um, I, I, I think, I, I, there are a few resource here, uh, Paul, uh, I, I can, I, I will share, uh, able to get that in, in one place, um, so you can, uh, get started as well. Uh, so you can register for the sandbox here to have your own website. And then you can create the extension. You can deploy, I think, for free, uh, on, um, on Azure or Google Cloud, whatever.
1:19:14 They have some, a free tier. You can deploy the front end there, and then you can, uh, uh, register the extension there, and then you can play around with that. Yeah.
1:19:24 Uh, I, I will say, uh, more, more about it, uh, in, uh, upcoming days. Yeah.
1:19:30 Chris McClellan: Any other questions from anyone? Everyone's been pretty quiet.
1:19:34 David Nguyen: Yeah. All right. Just-
1:19:36 Chris McClellan: So Simon's saying we better get to the slides
1:19:38 David Nguyen: …
1:19:39 more, more in Tableau.
1:19:40 Chris McClellan: So what's the best way to get the slides, David? Are you, are you gonna send them-
1:19:42 David Nguyen: Uh-
1:19:42 Chris McClellan: … through to me or, um, go to your website?
1:19:45 David Nguyen: Uh, yeah.
1:19:46 Uh, I can, I can, I can send it, uh, to you. Uh-
1:19:49 Chris McClellan: Yep
1:19:49 David Nguyen: … but, uh, uh, Chris. And then for, for, um, for the, um, sample code here and there, I, I will check it in later.
1:19:58 Um, and then I, I will put on my website there with, with a copy of the tutorial step by step. Yeah.
1:20:04 Chris McClellan: Yep.
1:20:05 David Nguyen: Uh, I will share with you the, the, the website, uh, the, the, the slide first.
1:20:09 Um, and we'll clean up the code, and then I will share more, like, tutorial copy so you can follow and then you can play around. Yeah.
1:20:16 Chris McClellan: Yeah. Best way to do it.
1:20:17 David Nguyen: Yeah.
1:20:20 Yeah.
1:20:22 Chris McClellan: And I know David had it on his slides, but there's, um, David's contact details on my slides.
1:20:28 And all-… it's, it's in the recording anyway, so you can come back and check that at any time.
1:20:33 David Nguyen: Yeah.
1:20:34 Chris McClellan: Um, now it's time for the survey.
1:20:37 Usually I joke that we're going to run the sur- the survey sort of runs as the presentation goes. But, uh, as I said, today David wrote all the questions for the quiz.
1:20:46 So I'm gonna start the quiz now at three minutes, and give everyone the full three minutes, you might need it sometimes-
1:20:53 David Nguyen: Yeah
1:20:54 Chris McClellan: … um, to go through and answer the questions, and, um, and see how we go today with the quiz.
1:21:01 David Nguyen: If, if it's not so boring, and then no one sleep-… and I think you should be able to answer all the question.
1:21:11 I think
1:21:11 Chris McClellan: Maybe this time people bugged down-
1:21:12 David Nguyen: Yeah. That's right
1:21:12 Chris McClellan: … and swapped the questions as well. I don't know.
1:21:14 David Nguyen: Yeah, yeah.
1:21:17 Chris McClellan: Okay. That's the end of time. I'll go through to… Oh, if I can press the buttons properly.
1:21:23 People can still be answering the quiz questions. I'll get up to them slowly.
1:21:27 Uh, our standard questions that we ask all the time, where do you live? State and country. We've got obviously a nice spread across Australia, India, and Sri Lanka today.
1:21:37 David Nguyen: Wow.
1:21:39 Chris McClellan: Pretty good. Didn't get to Europe, but it might be a little bit y- early for that. That's okay.
1:21:44 David Nguyen: Yeah, yeah.
1:21:46 Chris McClellan: Um, what industry do you work in? Just to see what people are using Tableau for. And a nice spread across there as well. IT, logistics, consulting, healthcare, pharmaceutical, higher ed, finance, and insurance.
1:22:01 And how many years have you been using Tableau? So a lot of people on the nine plus-
1:22:06 David Nguyen: Right
1:22:06 Chris McClellan: … end, and then it just spreads off easily, as it always does. As I said, this graph's always different because it just depends on, um, people's experience with Tableau.
1:22:18 Okay. The quiz questions for today is, what is the best way to cut the cost of AI insights on a busy dashboard?
1:22:27 The correct answer was to send aggregated data as a CSV and cache repeated requests and use a smaller model for simpler… Whatever the rest of that one was, because it got cut, got cut off.
1:22:41 Um, so most people got that right.
1:22:45 The next question was, where should the API key for your AI model live? Basically on the back end.
1:22:52 David Nguyen: On the back end. Yeah, on the back end.
1:22:53 Chris McClellan: At least so people can't see it because-
1:22:56 David Nguyen: Yeah.
1:22:56 Chris McClellan: … it's a little bit dangerous.
1:22:58 David Nguyen: Yeah.
1:23:00 Chris McClellan: Um, when an extension reads summary data from a worksheet, what does it see?
1:23:05 It sees what you can see, so any-
1:23:08 David Nguyen: What you can see
1:23:08 Chris McClellan: … any filters-
1:23:09 David Nguyen: Yeah
1:23:09 Chris McClellan: … and RLS is already applied.
1:23:12 David Nguyen: Yeah, yeah.
1:23:12 Chris McClellan: Which would be, um, hopefully right.
1:23:16 David Nguyen: Yeah.
1:23:16 Chris McClellan: And what must an extension call before it uses any other extensions API method? It has to do that initializeAsync function.
1:23:25 David Nguyen: Initialize. Yeah.
1:23:27 Chris McClellan: Because that basically sets up the handshake or something-
1:23:30 David Nguyen: Sets up the handset
1:23:30 Chris McClellan: … similar
1:23:30 David Nguyen: Yeah.
1:23:31 Chris McClellan: Yep.
1:23:31 David Nguyen: With the Tableau. Yeah.
1:23:35 Chris McClellan: And I think this is the final question. What is a Tableau dashboard extension? It's basically a TREX file or a T-REX file.
1:23:43 David Nguyen: Yeah. A TREX file or T-REX.
1:23:44 Chris McClellan: Um, which is that whole package of everything.
1:23:47 David Nguyen: Yeah.
1:23:47 Chris McClellan: Which is why they use a little dinosaur, as in T-Rex.
1:23:50 David Nguyen: Yeah.
1:23:50 Chris McClellan: Oh.
1:23:50 David Nguyen: Yeah, yeah.
1:23:53 Chris McClellan: A lot of people go, "What's a dinosaur got to do-
1:23:55 David Nguyen: Yeah
1:23:55 Chris McClellan: … with extensions?" And it's just like, no, it's a T-REX dinosaur-
1:23:58 David Nguyen: Yeah. That's it. Yeah
1:23:59 Chris McClellan: … because it's a TREX file.
1:24:02 So that's good.
1:24:03 David Nguyen: Yeah.
1:24:04 Chris McClellan: Um, what topic should be presented at the future TUG meeting? That must be the hardest question, because there was only two answers. So either more of this or dashboards.
1:24:14 And who would you like to see present?
1:24:17 Which is interesting because no one said anything.
1:24:22 There was one response, and I can't see that response. Okay. Interesting.
1:24:26 Okay. The leaderboard will be good because the leaderboard is David is on top because David-… knows the answers to David's own questions. So of course David took 14 seconds to get all the questions right.
1:24:38 I hope you got all the questions right.
1:24:42 Um, Paul, congratulations. Paul's-
1:24:44 David Nguyen: Oh
1:24:44 Chris McClellan: … really in first place, um, with 52.
1:24:47 David Nguyen: Wow.
1:24:48 Chris McClellan: I was second, Andy was third, and it trails off a little bit from there.
1:24:52 That's okay. If you didn't, if you didn't get all the answers right, you've learnt something along the way.
1:24:58 That quiz was brought to you by David who did the presentation today. Uh, anybody can submit questions for quizzes for either this TUG or any other TUG which uses that system.
1:25:09 Um, but you've seen both for Danny's presentation a few weeks ago and today, I actually ask the presenter if they want to submit questions.
1:25:18 David Nguyen: Yeah.
1:25:18 Chris McClellan: They don't have to at all.
1:25:20 Um, and David sent all those questions in, so it was really good.
1:25:23 Sort of helps me, but helps you, but helps everyone as well.
1:25:26 David Nguyen: Yeah, yeah, it did that
1:25:27 Chris McClellan: Yeah.
1:25:28 And that was good. Thank you for doing that.
1:25:30 David Nguyen: Yeah.
1:25:31 Chris McClellan: So the next TUG meeting is next week, early. I think that's a Tuesday, isn't it? October 20th for us-
1:25:37 David Nguyen: Yeah, yeah. October
1:25:38 Chris McClellan: … or 19th if you're in the US.
1:25:40 Um, Rob's talking about the classic topic that I hear a lot of times, hey, Tableau is not just for retail analytics.
1:25:49 Everyone thinks that Tableau is for retail stuff because it's only got Superstore in it. Uh, he's talking about-
1:25:54 David Nguyen: Yeah
1:25:55 Chris McClellan: … the US Sentencing Commission and how they looked at sentencing times and all that sort of, um, what? Not quite criminal stuff, but criminal history, judicial system type things as well.
1:26:09 So again, next week, that's the same… It, it's one meeting, it's just all those different times in the world depending on where you are, to give you a guide about, um, where you should be at that time.
1:26:22 Um, but this recording, everything that Team DataFam TUG does, a lot of other different TUGs, is on YouTube. You can also listen to the podcast on Spotify and Apple, or obviously email me at team.DataFam@Gmail.com if you've got any infor- if you've got any questions and information.
1:26:41 But, uh, again, thank you, David. As we said at the start… Well, I didn't say it, you said, um, this is first of a series of three. So about mid-November there's another one, and mid-December there's another one.
1:26:53 David Nguyen: Yeah.
1:26:53 Chris McClellan: We're talking about AI and Tableau extensions, all that other stuff as-
1:26:58 David Nguyen: Yes, that as well, uh, yes.
1:27:00 Chris McClellan: Yeah.
1:27:01 So, so thanks for your time today. That was awesome.
1:27:03 David Nguyen: Yeah. Thank, thank you, Chris. Thank you, everyone, for coming.
1:27:06 Chris McClellan: Easy.
1:27:07 David Nguyen: Yeah.
1:27:07 Chris McClellan: And if anyone wants to chat now, now's the time to unmute while we're here or, um, you know, it's Friday afternoon for me, or for most of us. Have a good weekend as well. We'll talk to you next week.
1:27:17 David Nguyen: Cool.
1:27:18 Chris McClellan: Thanks, Chris. Thanks, David. Okay. Thanks a lot.
1:27:20 David Nguyen: Yeah.
1:27:20 Chris McClellan: See you.
1:27:20 David Nguyen: Bye, Andy.
Shorts for this TUG
- Why build your own Tableau AI extension? 18:31
- Great dashboards still need someone to explain 9:02
- The AI found our top salesperson loses money 22:23
- When you DON'T need a custom AI extension 20:39
- Extensions only see the data you can see 37:07
- Send CSV, not JSON, to save AI tokens 41:28
- AI agents keep leaking API keys 50:42
- Lock down your API before someone burns your quota 1:01:25
- Three AI models, all correct, three different stories 58:38
- Output tokens cost 5x more than input 1:04:53
- Why companies are afraid to switch on Tableau AI 1:16:10
- Keep your AI data inside the country 1:14:46
- Why Tableau extensions have a T-Rex 1:23:37